Polymarket Traders Bet Against a US Ban on Chinese AI Models

Polymarket traders are pricing just a 15% chance that Washington cuts off public access to a Chinese AI model in 2026, down from 23% when the contract opened two weeks earlier. The number keeps falling even as insiders tell Axios that the Trump administration’s push to restrict models like Moonshot AI’s Kimi K3 is picking up steam again.

That gap between the odds and the rhetoric traces back to a technical fact no executive order can undo. Moonshot plans to publish Kimi K3’s full weights on July 27, and once a model’s parameters are sitting on the open internet, no single government can pull them back down.

Entity Lists and Procurement Threats

The Commerce Department considered adding several Chinese AI labs to its Entity List last year, a designation that would cut off their access to U.S. technology without a license, according to Axios reporting that surfaced the internal debate. The National Security Agency and the White House Office of the National Cyber Director also weighed a threat advisory on Chinese AI labs last year, and Commerce circulated draft rules using its supply chain authorities to target open-source models from China.

None of it went anywhere. Administration officials worried about strangling innovation killed each attempt. Then Kimi K3 landed on July 16, and the calculation changed.

Part of that shift is personnel. Sriram Krishnan, a former White House AI policy adviser who pushed back on restrictions, has since left, and national security hawks inside the administration have grown louder. A source familiar with the government discussions described the current approach bluntly: not a single sweeping ban, but a slower campaign of procurement rules, Entity List threats and public pressure aimed at the American companies already running Chinese models in production.

A White House official gave a similar answer to the Daily Caller News Foundation, saying the administration is committed to “promoting America’s open-source ecosystem and strengthening its security.” Neither the White House nor the Commerce Department responded to Axios’s request for comment on the specifics.

What Are Traders Betting On?

Polymarket’s contract asks whether the federal government will pass legislation, issue an executive order, impose an export control or take other formal action that generally cuts off U.S. public access to a major Chinese AI model by December 31. Right now that is priced at 15%, and a separate market on which lab runs the best model by the end of August shows how narrowly traders read Kimi K3’s win.

Polymarket’s own account flagged the story as it broke, posting that Trump officials are reportedly weighing blocking access to top Chinese models. Traders did not treat Kimi K3’s coding win as proof it would lead everywhere. They gave Anthropic a 91% shot at holding the overall crown into late August and gave Moonshot only 1%, even with the front-end benchmark trophy in hand.

Once the Weights Are Public, There Is No Recall

Kimi K3 is a 2.8 trillion parameter model, and Moonshot has said its Mixture-of-Experts design activates just 16 of its 896 experts for any given token, a setup the company says delivers roughly 2.5 times the scaling efficiency of the earlier Kimi K2. It runs a 1 million token context window, takes native image input and uses an attention mechanism Moonshot calls Kimi Delta Attention.

• 2.8 trillion parameters across a sparse Mixture-of-Experts architecture
• 16 of 896 experts activated per token, cutting compute cost sharply
• 1 million token context window with native image input
• Released July 16, with full open weights due July 27

Pricing is the detail that has pulled American companies toward it. Kimi K3 runs roughly 40% cheaper than flagship U.S. models on a per-token basis, with listings putting it near $3 per million input tokens and $15 per million output tokens. Self-hosting shifts hardware and power costs onto the company running it, but it also keeps sensitive data off a foreign API entirely, which is exactly why enterprises with heavy AI usage keep adopting it.

The full weight release was not sitting everywhere the day Kimi K3 launched. That is a real distinction, and it buys Washington a few more days. Once July 27 passes, the model can be mirrored and re-hosted by anyone, anywhere, and a ban built around a single switch has nothing left to flip.

Washington Already Ran This Play on DeepSeek

This is not the first time a Chinese model has forced this exact debate. DeepSeek’s rise in January last year triggered a similar shock, and the government response gives a preview of what “restricting access” has actually meant in practice.

1. January 2025: DeepSeek’s release rattles U.S. tech stocks and draws the first wave of Washington scrutiny.
2. February 7, 2025: Reps. Josh Gottheimer and Darin LaHood introduce the No DeepSeek on Government Devices Act.
3. March 3, 2025: The same lawmakers urge governors to ban DeepSeek from state devices; Virginia, Texas and New York do.
4. June 25, 2025: The bipartisan No Adversarial AI Act is introduced in the House and Senate.
5. July 20, 2026: Axios reports the same restriction debate reignited by Kimi K3, with the 2025 bill still in committee.

The No Adversarial AI Act would publish a public list of foreign adversary AI within 180 days of enactment and bar federal agencies from using it, with narrow exceptions for research. More than a year after it was introduced, it is still sitting with the House Oversight and Government Reform Committee. No nationwide public ban followed DeepSeek. What followed was device bans, three state-level restrictions and a bill that has not moved.

Sacks Calls Out a Duopoly Play

We are at a critical inflection point in AI policy. The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition.

David Sacks, the former White House AI and crypto czar who now co-chairs the President’s Council of Advisors on Science and Technology, wrote that on X. He followed it with a sharper line, saying rival labs “have laid their cards on the table” and that it was time for the rest of Silicon Valley to push back too.

Sacks had already flagged Kimi K3’s benchmark run as a warning sign, calling it “concerning” after the model ranked first on the Frontend Code Arena and scored near the frontier on other tests. CIA Director John Ratcliffe has taken the opposite lesson from the same data, arguing the U.S. cannot let China hold the advantage on frontier AI models. The White House statement to the Daily Caller lands somewhere between the two, framing the goal as strengthening security without naming a target.

Kimi K3 did not arrive in a vacuum. China’s Long March 10 rocket test for its moon program and a humanoid robot’s record setting 66 mile hike have added to a Washington sense that Chinese engineering is closing gaps faster than expected across more than one frontier field at once.

George Noble Puts a Number on the Risk

The policy fight is unfolding while a separate warning hangs over the entire AI trade. George Noble, a former Fidelity fund manager, said the fallout from an AI bubble bursting could run 17 times worse than the dot-com crash, which erased roughly $5 trillion from the Nasdaq.

“The fallout from this could really be much more significant,” Noble said, tying the risk to the scale of capital pouring into AI infrastructure regardless of who wins the model race. That backdrop raises the cost of getting this decision wrong in either direction, whether Washington restricts too little or moves so hard it strands American companies that already depend on cheap Chinese models to run their products.

Moonshot’s July 27 release date is the next real test of whether Washington can act before the weights are already everywhere.

Frequently Asked Questions

What does open-weight actually mean?

Open-weight means a company publishes the trained model’s parameters for anyone to download and run on their own hardware, without necessarily releasing the training code or data behind it. A fully open-source model goes further and publishes that pipeline too, which Moonshot has said it plans to do with Kimi K3 on July 27.

Has the U.S. already banned DeepSeek?

Not nationally. Several Commerce Department bureaus barred DeepSeek from staff devices, and Virginia, Texas and New York have banned it for state employees, but no state outside those three has followed and no federal law covers the general public.

What is the Commerce Department’s Entity List?

It is an export control tool that blocks listed companies from getting U.S. technology without a special license. Washington has used it before against firms including Huawei and TikTok owner ByteDance, and officials floated adding Chinese AI labs to it last year.

What happens if Kimi K3’s open weights go live as planned?

Independent developers get to verify Moonshot’s performance claims directly instead of relying on the company’s own benchmarks, and the model becomes freely copyable and re-hostable on servers outside any government’s reach, which is the scenario that makes a later ban far harder to enforce.

Disclaimer: This article is for informational purposes only and does not constitute financial or investment advice. Prediction market odds and AI market commentary involve real uncertainty, so consult a licensed professional before making investment decisions, and figures are accurate as of publication.